Optimal attribute search device, optimal attribute search method, and program
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Filing Date
- 2024-08-26
- Publication Date
- 2026-07-23
AI Technical Summary
Existing optimal attribute search devices suffer from reduced accuracy when the optimal value of an attribute falls outside a restricted range, as they fail to effectively update their search mechanisms.
The device includes an optimal value search unit, an observed value acquisition unit, and a function update unit that iteratively refine the acquisition function using observed values to enhance accuracy even when optimal values are outside the restricted range.
The device maintains high search accuracy by updating the acquisition function with observed values, ensuring accurate identification of optimal attributes even when they fall outside the predefined range.
Abstract
Description
Optimal attribute search device, optimal attribute search method and program
[0001] The present disclosure relates to an optimal attribute search device, an optimal attribute search method, and a program.
[0002] There is an optimal attribute search device that searches for an optimal value from among multiple values of an attribute of a certain object, which will maximize the effect of the treatment when the treatment is performed on the object. For example, Non-Patent Document 1 discloses an optimal attribute search device that calculates a probability distribution of the effect of a certain treatment when each of multiple values of the attribute of the certain object is given as a variable, and searches for an optimal value of the attribute based on the probability distribution.
[0003] Alaa, AM & van der Schaar, M. Bayesian inference of individualized treatment effects using multi-task gaussian processes. In Advances in Neural Information Processing Systems, pages 3424-3432, 2017.
[0004] The device disclosed in Non-Patent Document 1 has a problem in that when multiple values of an attribute given as a variable are restricted to fall within a certain range, if the optimal value of the attribute exists outside the certain range, the accuracy of searching for the optimal value may deteriorate.
[0005] The present disclosure has been made to solve the above-mentioned problems, and aims to provide an optimal attribute search device that can suppress deterioration in the search accuracy for optimal values even when multiple values for an attribute given as a variable are restricted to fall within a certain range and the optimal value of the attribute exists outside the certain range.
[0006] The optimal attribute search device according to the present disclosure includes: an optimal value search unit that, when each of a plurality of values of an attribute of a certain object is given as a variable, acquires a first acquisition function that outputs the difference between a treatment effect, which is the effect when a certain treatment is administered to the object, and an ineffective effect, which is the effect when the treatment is not administered to the object, and searches for one value from among the plurality of values of the attribute as a first optimal value based on the output of the first acquisition function; an observed value acquisition unit that acquires, as a first observed value, an observed value of the treatment effect corresponding to the first optimal value searched for by the optimal value search unit; and a function update unit that updates the first acquisition function using the first observed value acquired by the observed value acquisition unit. The optimal value search unit searches for the first optimal value based on the output of the first acquisition function after update by the function update unit, and outputs the searched first optimal value.
[0007] According to the present disclosure, when multiple values of an attribute given as a variable are restricted to fall within a certain range, even if the optimal value of the attribute exists outside the certain range, deterioration in the search accuracy for the optimal value can be suppressed.
[0008] FIG. 1 is a configuration diagram showing an optimal attribute search device according to a first embodiment. FIG. 2 is a hardware configuration diagram showing the hardware of the optimal attribute search device according to the first embodiment. FIG. 3 is a hardware configuration diagram of a computer when the optimal attribute search device is realized by software, firmware, or the like. FIG. 4 is a flowchart showing an optimal attribute search method, which is a processing procedure of the optimal attribute search device. ０，ｉ is an explanatory diagram showing the difference between the effect of not vaccinating and the effect of vaccinating when X is the age, the treatment for the subject is vaccination, and the effect is the infection rate of the disease. ０，ｉ is an explanatory diagram showing the difference between the effect of not vaccinating and the effect of vaccinating when X is the age, the treatment for the subject is vaccination, and the effect is the infection rate of the disease. ０，ｉFIG. 10A is an explanatory diagram showing a causal graph when the attribute directly related to the infection rate is gender, FIG. 10B is an explanatory diagram showing a causal graph when the attribute directly related to the infection rate is age, FIG. 10C is an explanatory diagram showing a causal graph when the attributes directly related to the infection rate are both gender and age, and FIG. 10D is an explanatory diagram showing a causal graph when the attribute directly related to the infection rate is cholesterol level.
[0009] In order to explain the present disclosure in more detail, embodiments of the present disclosure will be described below with reference to the accompanying drawings.
[0010] Embodiment 1. Fig. 1 is a configuration diagram showing an optimal attribute search device according to embodiment 1. Fig. 2 is a hardware configuration diagram showing the hardware of the optimal attribute search device according to embodiment 1. The optimal attribute search device shown in Fig. 1 includes a data storage unit 1, an optimal value search unit 2, an observed value acquisition unit 3, and a function update unit 4.
[0011] In the optimal attribute search device shown in FIG. 1 , when multiple values of an attribute of a certain object are given as variables, observed values of a treatment effect, which is the effect when a certain treatment is administered to the object, and observed values of an ineffective treatment effect, which is the effect when a certain treatment is not administered to the object, are obtained. The attributes of a certain object include, for example, the age, gender, or cholesterol level of the object. A certain treatment includes, for example, administering a vaccine to a subject or distributing coupons to a subject. If the certain treatment is administering a vaccine to a subject, the observed values of the treatment effect and the ineffective treatment effect are, for example, the infection rate of the disease corresponding to the vaccine. If the certain treatment is distributing coupons to subjects, the observed values of the treatment effect and the ineffective treatment effect are, for example, the purchase rate of the product covered by the coupon.
[0012] The data storage unit 1 is realized by, for example, a data storage circuit 11 shown in FIG. ０ The observation data D ０ is an attribute X of an object ０，ｉ and attribute X ０，ｉ The observed value of no treatment effect y corresponding to ０ ０，ｉ and an attribute X of a certain object ０，ｉ and attribute X ０，ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ It contains pairs with .
[0013] The optimum value search unit 2 is realized by, for example, an optimum value search circuit 12 shown in FIG. ０，ｉ When each of the multiple values in is given as a variable, a first acquisition function is obtained that outputs the difference between the treatment effect when a certain treatment is administered to the subject and the no-treatment effect when the treatment is not administered to the subject. The first acquisition function in the initial stage is obtained by using the observation data D stored in the data storage unit 1. ０ The calculation is based on the observation data D ０ Since the process of calculating the first acquisition function based on the observation data D is a known technique, a detailed description thereof will be omitted. If the first acquisition function is stored in the data storage unit 1, the optimum value search unit 2 acquires the first acquisition function from the data storage unit 1. However, this is only an example, and the optimum value search unit 2 may acquire the first acquisition function from outside the optimum attribute search device shown in FIG. 1. In addition, when the optimum value search unit 2 calculates the first acquisition function based on the observation data D, ０ The optimum value search unit 2 may calculate the first acquisition function based on the output of the first acquisition function. ０，ｉ Among the multiple values in Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ to the observation value acquisition unit 3, and outputs the first acquisition function to the function update unit 4.
[0014] The observation value acquisition unit 3 is realized by, for example, the observation value acquisition circuit 13 shown in FIG. 2. The observation value acquisition unit 3 receives the first optimum value x Ｉ The observation value acquisition unit 3 acquires the first optimal value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ The observation value acquisition unit 3 acquires the first observation value y １ ０，ｉ is output to the function update unit 4.
[0015] The function update unit 4 is realized by, for example, the function update circuit 14 shown in Fig. 2. The function update unit 4 acquires the first acquisition function from the optimum value search unit 2. The function update unit 4 acquires the first observation value y １ ０，ｉ The function update unit 4 obtains the first observed value y １ ０，ｉ The function update unit 4 updates the first acquisition function using the above formula. The function update unit 4 outputs the updated first acquisition function to the optimum value search unit 2.
[0016] The optimum value search unit 2 searches for a first optimum value based on the output of the first acquisition function after update by the function update unit 4, and outputs the searched first optimum value, for example, to the outside, and also outputs the difference output from the updated first acquisition function, for example, to the outside.
[0017] 1, it is assumed that each of the components of the optimum attribute search device, that is, the data storage unit 1, the optimum value search unit 2, the observed value acquisition unit 3, and the function update unit 4, is realized by dedicated hardware as shown in Fig. 2. That is, it is assumed that the optimum attribute search device is realized by a data storage circuit 11, an optimum value search circuit 12, an observed value acquisition circuit 13, and a function update circuit 14. Here, the data storage circuit 11 corresponds to, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory) or a flash memory, a magnetic disk, a flexible disk, an optical disk, a compact disk, a minidisk, or a DVD (Digital Versatile Disc). Furthermore, each of the optimum value search circuit 12, the observation value acquisition circuit 13, and the function update circuit 14 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.
[0018] The components of the optimal attribute search device are not limited to those realized by dedicated hardware, and the optimal attribute search device may be realized by software, firmware, or a combination of software and firmware. Software or firmware is stored in the memory of a computer as a program. A computer refers to hardware that executes a program, and includes, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a central processing unit, a processing unit, an arithmetic unit, a microprocessor, a microcomputer, a processor, or a DSP (Digital Signal Processor).
[0019] 3 is a hardware configuration diagram of a computer when the optimal attribute search device is realized by software, firmware, or the like. When the optimal attribute search device is realized by software, firmware, or the like, the data storage unit 1 is configured on the memory 21 of the computer. Programs for causing the computer to execute the respective processing procedures of the optimal value search unit 2, the observed value acquisition unit 3, and the function update unit 4 are stored in the memory 21. Then, a processor 22 of the computer executes the programs stored in the memory 21.
[0020] 2 shows an example in which each of the components of the optimal attribute search device is realized by dedicated hardware, and Fig. 3 shows an example in which the optimal attribute search device is realized by software, firmware, etc. However, this is merely an example, and some of the components in the optimal attribute search device may be realized by dedicated hardware, and the remaining components may be realized by software, firmware, etc.
[0021] Next, the operation of the optimum attribute search device shown in Fig. 1 will be described. Fig. 4 is a flowchart showing an optimum attribute search method, which is a processing procedure of the optimum attribute search device. Observation data D stored in the data storage unit 1 ０ is an attribute X of a certain object as shown in the following formula (1). ０，ｉ and attribute X ０，ｉ The observed value of no treatment effect y corresponding to ０ ０，ｉ The pair {y ０ ０，ｉ , X ０，ｉ} and an attribute X of an object ０，ｉ and attribute X ０，ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ The pair {y １ ０，ｉ , X ０，ｉ}.
[0022]
[0023] Figure 5 shows the relationship between the target and the attribute X of the target. ０，ｉis an explanatory diagram showing the difference between the effect of not vaccinated and the effect of vaccinated when age is the target, the treatment is vaccination, and the effect is the infection rate of the disease. In Figure 5, △ is the observed value y of the no-treatment effect, which is the effect of not vaccinated. ０ ０，ｉ , and ● indicates the observed value of the treatment effect y １ ０，ｉ The gray area including the dashed line connecting multiple triangles represents the observed value y ０ ０，ｉ The probability distribution P(y ０ The gray area including the dashed line connecting multiple ● indicates the observed value y of the treatment effect. １ ０，ｉ The probability distribution P(y １ hat|X). In the text of the specification, y ０ , y １ Since the symbol "^" cannot be added above y ０ Hat, Y １ It is written as a hat.
[0024] If the first acquisition function is stored in the data storage unit 1, the optimum value search unit 2 acquires the first acquisition function from the data storage unit 1. The first acquisition function is a function of an attribute X of a certain object. ０，ｉ In the example of FIG. 5, the first acquisition function is a function that outputs the difference between the treatment effect and the no-treatment effect when each of the multiple values in attribute X is given as a variable. ０，ｉ When each value in is given as a variable, it is a function that outputs the difference between the dashed line connecting multiple triangles and the dashed line connecting multiple black dots. In the example of FIG. 5, the observed value of the triangle mark is ０，ｉ The observed value of the no-treatment effect y ０ ０，ｉ The probability distribution P(y ０ The error of attribute X is large. ０，ｉThe observed value of the treatment effect y １ ０，ｉ The probability distribution P(y １ hat |X) has a large error.
[0025] The optimum value search unit 2 calculates the attribute X based on the output of the first acquisition function. ０，ｉ Among the multiple values in Ｉ (Step ST1 in FIG. 4). Specifically, the optimum value search unit 2 searches for x Ｉ = {x Ｉ，ｉ} ｉ ＮＩ Monte Carlo sampling is performed, and the attribute X that maximizes the difference between the treatment effect and the no-treatment effect is calculated as shown in the following formula (2). ０，ｉ to the first optimum value x Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ to the observation value acquisition unit 3, and outputs the first acquisition function to the function update unit 4.
[0026]
[0027] The observation value acquisition unit 3 receives the first optimum value x from the optimum value search unit 2. Ｉ The observation value acquisition unit 3 acquires the first optimal value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ is obtained as the first observation value (step ST2 in FIG. 4). Ｉ The first observation y １ ０，ｉ The process of obtaining the first optimal value x is not limited to the Bayesian optimization method, and any method may be used. Ｉ The first observed value of the treatment effect, y １ ０，ｉ The observation value acquisition unit 3 acquires the first observation value y １ ０，ｉ is output to the function update unit 4.
[0028] The function update unit 4 acquires the first acquisition function from the optimum value search unit 2. The function update unit 4 acquires the first observation value y １ ０，ｉ The function update unit 4 obtains the first observed value y １ ０，ｉ (Step ST3 in FIG. 4). Specifically, the function update unit 4 updates the first acquisition function by using not only the observed values marked with ● but also the observed values marked with ◯ in the younger age range, to obtain the observed value y １ ０，ｉ The probability distribution P(y １ Then, the function update unit 4 calculates the observed value y ０ ０，ｉ The probability distribution P(y ０ hat |X) and the calculated observed value of the treatment effect y １ ０，ｉ The probability distribution P(y １ The first acquisition function is calculated from the probability distribution P(y １ The calculation process for each of the first acquisition function and the first acquisition function is well known, and therefore a detailed description thereof will be omitted. １ ０，ｉ The probability distribution P(y １ hat |X), the accuracy of the calculated first acquisition function is improved compared to the accuracy of the first acquisition function in the initial stage. The function update unit 4 outputs the updated first acquisition function to the optimum value search unit 2.
[0029] The optimum value search unit 2 obtains the updated first acquisition function from the function update unit 4. The optimum value search unit 2 obtains the first optimum value x Ｉis given to the updated first acquisition function, the difference output from the first acquisition function is compared with the target value. The target value is a preset threshold value, and may be stored in the internal memory of the optimum value search unit 2, or may be given from outside the optimum attribute search device shown in FIG. 1. If the difference output from the first acquisition function is smaller than the target value (step ST4 in FIG. 4: NO), the optimum value search unit 2 calculates the attribute X based on the output of the updated first acquisition function. ０，ｉ Among the multiple values in Ｉ (Step ST5 in FIG. 4). The first optimum value x Ｉ The search process is carried out by determining the first optimum value x Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ is output to the observation value acquisition unit 3.
[0030] The observation value acquisition unit 3 receives the first optimum value x from the optimum value search unit 2. Ｉ The observation value acquisition unit 3 acquires the first optimal value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ The observation value acquisition unit 3 acquires the first observation value y １ ０，ｉ is output to the function update unit 4.
[0031] The function update unit 4 receives the first observed value y １ ０，ｉ The function update unit 4 obtains the first observed value y １ ０，ｉ The updated first acquisition function is further updated using (step ST3 in FIG. 4). As long as the difference output from the first acquisition function is smaller than the target value, the processes of steps ST2 to ST5 are repeated.
[0032] If the difference output from the first acquisition function is equal to or greater than the target value (step ST4 in FIG. 4: YES), the optimum value search unit 2 calculates the attribute X based on the output of the updated first acquisition function. ０，ｉAmong the multiple values in Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ The optimum value search unit 2 outputs the first optimum value x Ｉ is given to the updated first acquisition function, the difference output from the updated first acquisition function is output, for example, to the outside.
[0033] In the first embodiment described above, the optimal attribute search device is configured to include: an optimal value search unit 2 that, when multiple values of an attribute of a certain object are given as variables, acquires a first acquisition function that outputs the difference between a treatment effect, which is the effect when a certain treatment is applied to the object, and an ineffective treatment effect, which is the effect when the treatment is not applied to the object; and searches for one of the multiple values of the attribute as a first optimal value based on the output of the first acquisition function; an observed value acquisition unit 3 that acquires, as a first observed value, an observed value of the treatment effect corresponding to the first optimal value searched by the optimal value search unit 2; and a function update unit 4 that updates the first acquisition function using the first observed value acquired by the observed value acquisition unit 3. Furthermore, the optimal value search unit 2 searches for the first optimal value based on the output of the first acquisition function after update by the function update unit 4, and outputs the searched first optimal value. Therefore, when multiple values of an attribute given as variables are restricted to fall within a certain range, the optimal attribute search device can suppress deterioration in the accuracy of the search for the optimal value even when the optimal value of the attribute is outside the certain range.
[0034] Second Embodiment In a second embodiment, an optimum attribute search device will be described in which the function update unit 4 updates the first acquisition function using the second observed value.
[0035] The configuration of the optimal attribute search device according to the second embodiment is the same as that of the optimal attribute search device according to the first embodiment. Therefore, the configuration diagram showing the optimal attribute search device according to the second embodiment is shown in FIG. 1. In the second embodiment, the optimal value search unit 2 uses the target attribute X in addition to the first acquisition function. ０，ｉWhen each of the multiple values in x is given as a variable, a second acquisition function is obtained that outputs the variance of the treatment effect. The optimal value search unit 2 obtains the second acquisition function based on the output of the second acquisition function. ０，ｉ Among the multiple values in Ｉ and the second optimum value x Ｉ to the observation value acquisition unit 3. The optimum value search unit 2 outputs the first acquisition function and the second acquisition function to the function update unit 4.
[0036] The observation value acquisition unit 3 acquires the second optimum value x searched for by the optimum value search unit 2 instead of acquiring the first observation value. Ｉ The observed value of the treatment effect corresponding to the second observation y １ ０，ｉ The observation value acquisition unit 3 acquires the second observation value y １ ０，ｉ to the function update unit 4. The function update unit 4 outputs the first observed value y １ ０，ｉ Instead of the second observation value y １ ０，ｉ The function update unit 4 updates the first acquisition function using the second observation value y １ ０，ｉ Update the second acquisition function using:
[0037] Next, the operation of the optimum attribute search device according to the second embodiment will be described. ０，ｉ is an explanatory diagram showing the difference between the effect of not vaccinated and the effect of vaccinated when age is the target, the treatment is vaccination, and the effect is the infection rate of the disease. In Figure 6, △ is the observed value y ０ ０，ｉ , and ● indicates the observed value of the treatment effect y １ ０，ｉ The gray area including the dashed line connecting multiple triangles represents the observed value y ０ ０，ｉ The probability distribution P(y ０The gray area including the dashed line connecting multiple ● indicates the observed value y of the treatment effect. １ ０，ｉ The probability distribution P(y １ hat |X).
[0038] If the first acquisition function is stored in the data storage unit 1, the optimum value search unit 2 acquires the first acquisition function from the data storage unit 1. If the second acquisition function is stored in the data storage unit 1, the optimum value search unit 2 acquires the second acquisition function from the data storage unit 1. The second acquisition function is a function of an attribute X of a certain object. ０，ｉ is a function that outputs the variance of the treatment effect when each of the multiple values in is given as a variable. In the example of Figure 6, the gray area surrounding the dashed line connecting the multiple ●s indicates the variance of the treatment effect.
[0039] The optimum value search unit 2 calculates the attribute X based on the output of the second acquisition function. ０，ｉ Among the multiple values in Ｉ Specifically, the optimum value search unit 2 searches for x Ｉ = {x Ｉ，ｉ} ｉ ＮＩ is sampled by Monte Carlo, and the attribute X ０，ｉ to the second optimum value x Ｉ The optimum value search unit 2 searches for the second optimum value x Ｉ to the observation value acquisition unit 3, and outputs the first acquisition function and the second acquisition function to the function update unit 4.
[0040]
[0041] The observation value acquisition unit 3 receives the second optimum value x from the optimum value search unit 2. Ｉ The observation value acquisition unit 3 acquires the second optimal value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ is obtained as the second observation. The second optimal value x Ｉ The second observation y １０，ｉ The process of obtaining the second optimal value x is not limited to the Bayesian optimization method, and any method may be used. Ｉ The second observation of the treatment effect, y, corresponds to １ ０，ｉ The observation value acquisition unit 3 acquires the second observation value y １ ０，ｉ is output to the function update unit 4.
[0042] The function update unit 4 acquires the first acquisition function and the second acquisition function from the optimum value search unit 2. The function update unit 4 acquires the second observation value y １ ０，ｉ The function update unit 4 obtains the second observed value y １ ０，ｉ Specifically, the function update unit 4 updates the first acquisition function by using not only the observed values marked with ● but also the observed values marked with ◯ in the younger age range, and calculates the observed value y １ ０，ｉ The probability distribution P(y １ Then, the function update unit 4 calculates the observed value y ０ ０，ｉ The probability distribution P(y ０ hat |X) and the calculated observed value of the treatment effect y １ ０，ｉ The probability distribution P(y １ The function update unit 4 calculates a first acquisition function from the calculated observed value y |X of the treatment effect. １ ０，ｉ The probability distribution P(y １ The second acquisition function is calculated from the second acquisition function (X). The calculation process of the second acquisition function itself is a known technique, so a detailed explanation will be omitted. The observed values with the marks ● as well as the observed values with the marks ○ are used to calculate the observed value y １ ０，ｉ The probability distribution P(y １hat |X), the accuracy of the calculated first acquisition function is improved compared to the accuracy of the first acquisition function in the initial stage, and the accuracy of the calculated second acquisition function is improved compared to the accuracy of the second acquisition function in the initial stage. The function update unit 4 outputs the updated first acquisition function and the updated second acquisition function to the optimum value search unit 2.
[0043] The optimum value search unit 2 obtains the updated first acquisition function and the updated second acquisition function from the function update unit 4. The optimum value search unit 2 obtains the second optimum value x Ｉ When the updated first acquisition function is given a difference, the difference output from the first acquisition function is compared with the target value. If the difference output from the first acquisition function is smaller than the target value, the optimum value search unit 2 calculates the attribute X based on the output of the updated second acquisition function. ０，ｉ Among the multiple values in Ｉ The optimum value search unit 2 searches for the second optimum value x Ｉ is output to the observation value acquisition unit 3.
[0044] The observation value acquisition unit 3 receives the second optimum value x from the optimum value search unit 2. Ｉ The observation value acquisition unit 3 acquires the second optimal value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ The observation value acquisition unit 3 acquires the second observation value y １ ０，ｉ is output to the function update unit 4.
[0045] The function update unit 4 receives the second observed value y １ ０，ｉ The function update unit 4 obtains the second observed value y １ ０，ｉ The function update unit 4 further updates the updated first acquisition function using the second observation value y １ ０，ｉThe updated second acquisition function is further updated using the following formula: As long as the difference output from the first acquisition function is smaller than the target value, the process is repeated in the same manner as in the optimal attribute search device according to the first embodiment.
[0046] If the difference output from the first acquisition function is equal to or greater than the target value, the optimum value search unit 2 calculates the attribute X based on the output of the updated first acquisition function. ０，ｉ Among the multiple values in Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ The optimum value search unit 2 outputs the first optimum value x Ｉ is given to the updated first acquisition function, the difference output from the updated first acquisition function is output, for example, to the outside.
[0047] In the second embodiment described above, the optimal value search unit 2 acquires, in addition to the first acquisition function, a second acquisition function that outputs the variance of the treatment effect when each of multiple values in the target attribute is given as a variable, and searches for one value from among the multiple values in the attribute as a second optimal value based on the output of the second acquisition function. The observed value acquisition unit 3 acquires, as a second observed value, an observed value of the treatment effect corresponding to the second optimal value searched for by the optimal value search unit 2, instead of acquiring the first observed value. The optimal attribute search device is configured so that the function update unit 4 updates the first acquisition function using the second observed value acquired by the observed value acquisition unit 3, instead of the first observed value. The optimal value search unit 2 also searches for a first optimal value based on the output of the first acquisition function after update by the function update unit 4, and outputs the searched first optimal value. Therefore, when multiple values of an attribute given as a variable are restricted to fall within a certain range, the optimum attribute search device can suppress deterioration in the search accuracy for the optimum value even when the optimum value of the attribute is outside the certain range.
[0048] Third Embodiment In the third embodiment, an optimal attribute search device will be described in which the function update unit 4 updates the first acquisition function using the first observed value and the second observed value. The configuration of the optimal attribute search device according to the third embodiment is the same as that of the optimal attribute search device according to the first embodiment. Therefore, the configuration diagram showing the optimal attribute search device according to the third embodiment is FIG. 1. In the third embodiment, the optimal value search unit 2 acquires a second acquisition function in addition to the first acquisition function. The optimal value search unit 2 updates the attribute X based on the output of the first acquisition function. ０，ｉ Among the multiple values in Ｉ and based on the output of the second acquisition function, ０，ｉ Among the multiple values in Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ and the second optimum value x Ｉ and outputs the first acquisition function and the second acquisition function to the function updater 4.
[0049] The observation value acquisition unit 3 obtains the first optimum value x Ｉ The observed value of the treatment effect corresponding to the first observation y １ ０，ｉ and the second optimum value x Ｉ The observed value of the treatment effect corresponding to the second observation y １ ０，ｉ The observation value acquisition unit 3 acquires the first observation value y １ ０，ｉ and the second observation y １ ０，ｉ and is output to the function update unit 4. The function update unit 4 outputs the first observed value y １ ０，ｉ and the second observation y １ ０，ｉ The function update unit 4 updates the first acquisition function using the first observation value y １ ０，ｉ and the second observation y １ ０，ｉ Update the second acquisition function using:
[0050] Next, the operation of the optimum attribute search device according to the second embodiment will be described. ０，ｉ is an explanatory diagram showing the difference between the effect of not vaccinated and the effect of vaccinated when age is the target, the treatment is vaccination, and the effect is the infection rate of the disease. In Figure 7, △ is the observed value y ０ ０，ｉ , and ● indicates the observed value of the treatment effect y １ ０，ｉ The gray area including the dashed line connecting multiple triangles represents the observed value y ０ ０，ｉ The probability distribution P(y ０ The gray area including the dashed line connecting multiple ● indicates the observed value y of the treatment effect. １ ０，ｉ The probability distribution P(y １ hat |X).
[0051] If the first acquisition function is stored in the data storage unit 1, the optimum value search unit 2 acquires the first acquisition function from the data storage unit 1. If the second acquisition function is stored in the data storage unit 1, the optimum value search unit 2 acquires the second acquisition function from the data storage unit 1.
[0052] The optimum value search unit 2 calculates the attribute X based on the output of the first acquisition function. ０，ｉ Among the multiple values in Ｉ Furthermore, the optimum value search unit 2 searches for the attribute X based on the output of the second acquisition function. ０，ｉ Among the multiple values in Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ and the second optimum value x Ｉ and outputs the first acquisition function and the second acquisition function to the function updater 4.
[0053] The observation value acquisition unit 3 receives the first optimum value x from the optimum value search unit 2. Ｉ and the second optimum value xＩ The observation value acquisition unit 3 acquires the first optimal value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ The observation value acquiring unit 3 acquires the second optimum value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ The observation value acquisition unit 3 acquires the first observation value y １ ０，ｉ and the second observation y １ ０，ｉ and are output to the function update unit 4.
[0054] The function update unit 4 acquires the first acquisition function and the second acquisition function from the optimum value search unit 2. The function update unit 4 acquires the first observation value y １ ０，ｉ and the second observation y １ ０，ｉ The function update unit 4 obtains the first observed value y １ ０，ｉ and the second observation y １ ０，ｉ Specifically, the function update unit 4 updates the first acquisition function by using not only the observed values marked with ● but also the observed values marked with ◯ in the younger age range, to obtain the observed value y １ ０，ｉ The probability distribution P(y １ Then, the function update unit 4 calculates the observed value y ０ ０，ｉ The probability distribution P(y ０ hat |X) and the calculated observed value of the treatment effect y １ ０，ｉ The probability distribution P(y １ The function update unit 4 calculates a first acquisition function from the calculated observed value y |X of the treatment effect. １ ０，ｉ The probability distribution P(y １The function update unit 4 outputs the updated first acquisition function and the updated second acquisition function to the optimum value search unit 2.
[0055] The optimum value search unit 2 obtains the updated first acquisition function and the updated second acquisition function from the function update unit 4. The optimum value search unit 2 obtains the first optimum value x Ｉ When the first acquisition function after update is given, the difference output from the first acquisition function is compared with the target value. If the difference output from the first acquisition function is smaller than the target value, the optimum value search unit 2 calculates the attribute X based on the output of the first acquisition function after update. ０，ｉ Among the multiple values in Ｉ Furthermore, the optimum value search unit 2 searches for the attribute X based on the output of the second acquisition function. ０，ｉ Among the multiple values in Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ and the second optimum value x Ｉ and are output to the observation value acquisition unit 3.
[0056] The observation value acquisition unit 3 receives the first optimum value x from the optimum value search unit 2. Ｉ and the second optimum value x Ｉ The observation value acquisition unit 3 acquires the first optimal value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ The observation value acquiring unit 3 acquires the second optimum value x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ ０，ｉ The observation value acquisition unit 3 acquires the first observation value y １ ０，ｉ and the second observation y １ ０，ｉ and are output to the function update unit 4.
[0057] The function update unit 4 receives the first observed value y １０，ｉ and the second observation y １ ０，ｉ The function update unit 4 obtains the first observed value y １ ０，ｉ and the second observation y １ ０，ｉ The function update unit 4 further updates the updated first acquisition function using the first observation value y １ ０，ｉ and the second observation y １ ０，ｉ The updated second acquisition function is further updated using the above. As long as the difference output from the first acquisition function is smaller than the target value, the process is repeated in the same manner as in the optimal attribute search device according to the first embodiment.
[0058] If the difference output from the first acquisition function is equal to or greater than the target value, the optimum value search unit 2 calculates the attribute X based on the output of the updated first acquisition function. ０，ｉ Among the multiple values in Ｉ The optimum value search unit 2 searches for the first optimum value x Ｉ The optimum value search unit 2 outputs the first optimum value x Ｉ is given to the updated first acquisition function, the difference output from the updated first acquisition function is output, for example, to the outside.
[0059] In the third embodiment described above, the optimal value search unit 2 acquires, in addition to the first acquisition function, a second acquisition function that outputs the variance of the treatment effect when each of multiple values of the target attribute is given as a variable. Based on the output of the second acquisition function, the optimal value search unit 2 searches for one of the multiple values of the attribute as a second optimal value. The observed value acquisition unit 3 acquires, in addition to the first observed value, an observed value of the treatment effect corresponding to the second optimal value searched for by the optimal value search unit 2 as a second observed value. The optimal attribute search device is configured so that the function update unit 4 updates the first acquisition function using the first observed value and the second observed value. The optimal value search unit 2 also searches for the first optimal value based on the output of the first acquisition function updated by the function update unit 4, and outputs the searched first optimal value. Therefore, the optimal attribute search device can suppress deterioration in the accuracy of the search for the optimal value even when multiple values of the attribute given as variables are restricted to fall within a certain range and the optimal value of the attribute is outside the certain range. Furthermore, the optimum attribute search device has improved accuracy in searching for optimum values of attributes compared to the optimum attribute search devices according to the first and second embodiments.
[0060] Fourth Embodiment In a fourth embodiment, an optimum attribute search device will be described that displays a causal graph showing the causal relationships between multiple types of attributes of a target and treatment effects when the target has multiple types of attributes.
[0061] FIG. 8 is a configuration diagram showing an optimal attribute search device according to embodiment 4. In FIG. 8, the same reference numerals as in FIG. 1 indicate the same or corresponding parts, and detailed description thereof will be omitted. FIG. 9 is a hardware configuration diagram showing the hardware of an optimal attribute search device according to embodiment 4. In FIG. 9, the same reference numerals as in FIG. 2 indicate the same or corresponding parts, and detailed description thereof will be omitted. The optimal attribute search device shown in FIG. 8 includes a data storage unit 1, an optimal value search unit 2, an observed value acquisition unit 3, a function update unit 4, a causal graph acquisition unit 5, a likelihood calculation unit 6, a causal graph selection unit 7, and a display processing unit 8.
[0062] The causal graph acquisition unit 5 is realized by, for example, a causal graph acquisition circuit 15 shown in Fig. 9. The causal graph acquisition unit 5 acquires a plurality of causal graphs that show different causal relationships between a plurality of types of attributes and treatment effects. The causal graph acquisition unit 5 outputs the plurality of causal graphs to the likelihood calculation unit 6 and the causal graph selection unit 7, respectively.
[0063] The likelihood calculation unit 6 is realized by, for example, a likelihood calculation circuit 16 shown in FIG. 9. The likelihood calculation unit 6 receives the observed data D ０ and obtains a first observation value relating to each type of attribute. The likelihood calculation unit 6 obtains a plurality of causal graphs from the causal graph acquisition unit 5. The likelihood calculation unit 6 obtains the observed data D ０ and the first observation value for each attribute, the likelihood calculation unit 6 calculates the likelihood of each causal graph. The likelihood calculation unit 6 outputs the likelihood of each causal graph to the causal graph selection unit 7.
[0064] The causal graph selection unit 7 is realized by, for example, a causal graph selection circuit 17 shown in Fig. 9. The causal graph selection unit 7 acquires a plurality of causal graphs from the causal graph acquisition unit 5, and acquires the likelihood of each causal graph from the likelihood calculation unit 6. The causal graph selection unit 7 selects one of the plurality of causal graphs based on the likelihoods of the plurality of causal graphs calculated by the likelihood calculation unit 6. The causal graph selection unit 7 outputs the selected causal graph to the display processing unit 8.
[0065] The display processing unit 8 is realized by, for example, a display processing circuit 18 shown in Fig. 9. The display processing unit 8 acquires the selected causal graph from the causal graph selection unit 7. The display processing unit 8 displays the selected causal graph on, for example, a display (not shown).
[0066] 8, it is assumed that each of the components of the optimal attribute search device, namely, a data storage unit 1, an optimal value search unit 2, an observed value acquisition unit 3, a function update unit 4, a causal graph acquisition unit 5, a likelihood calculation unit 6, a causal graph selection unit 7, and a display processing unit 8, is realized by dedicated hardware such as that shown in FIG. 9. That is, it is assumed that the optimal attribute search device is realized by a data storage circuit 11, an optimal value search circuit 12, an observed value acquisition circuit 13, a function update circuit 14, a causal graph acquisition circuit 15, a likelihood calculation circuit 16, a causal graph selection circuit 17, and a display processing circuit 18. Each of the optimal value search circuit 12, the observed value acquisition circuit 13, the function update circuit 14, the causal graph acquisition circuit 15, the likelihood calculation circuit 16, the causal graph selection circuit 17, and the display processing circuit 18 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof.
[0067] The components of the optimal attribute search device are not limited to those realized by dedicated hardware, and the optimal attribute search device may be realized by software, firmware, or a combination of software and firmware. When the optimal attribute search device is realized by software, firmware, or the like, the data storage unit 1 is configured on a memory 21 shown in FIG. 3. Programs for causing a computer to execute the respective processing procedures of the optimal value search unit 2, the observed value acquisition unit 3, the function update unit 4, the causal graph acquisition unit 5, the likelihood calculation unit 6, the causal graph selection unit 7, and the display processing unit 8 are stored in the memory 21. Then, a processor 22 shown in FIG. 3 executes the programs stored in the memory 21.
[0068] 9 shows an example in which each of the components of the optimal attribute search device is realized by dedicated hardware, while Fig. 3 shows an example in which the optimal attribute search device is realized by software, firmware, etc. However, this is merely an example, and some of the components in the optimal attribute search device may be realized by dedicated hardware, and the remaining components may be realized by software, firmware, etc.
[0069] Next, the operation of the optimal attribute search device shown in Fig. 8 will be described. Since the components other than the causal graph acquisition unit 5, likelihood calculation unit 6, causal graph selection unit 7, and display processing unit 8 are generally the same as those of the optimal attribute search device shown in Fig. 1, the operations of the causal graph acquisition unit 5, likelihood calculation unit 6, causal graph selection unit 7, and display processing unit 8 will mainly be described here.
[0070] When there are multiple types of attributes of a target, for example, if a certain treatment is vaccination of a target, the attributes of each type may include, for example, age, sex, and cholesterol level. When there are multiple types of attributes of a target, the optimum value search unit 2 searches for each type of attribute X based on the output of the first acquisition function. Ｉ，ｉ Among the multiple values in Ｉ The optimum value search unit 2 searches for each type of attribute X Ｉ，ｉ The first optimum value x Ｉ to the observation value acquisition unit 3, and outputs the first acquisition function to the function update unit 4.
[0071] The observation value acquisition unit 3 receives the attribute X of each type from the optimum value search unit 2. Ｉ，ｉ The first optimum value x Ｉ The observation value acquisition unit 3 acquires the first optimal values x by, for example, performing a Bayesian optimization technique. Ｉ The observed value of the treatment effect corresponding to y １ Ｉ，ｉ The observation value acquisition unit 3 acquires the observation data D stored in the data storage unit 1 as the first observation value. ０ and each type of attribute X Ｉ，ｉ The first observation y １ Ｉ，ｉ and are output to the likelihood calculation unit 6.
[0072] As shown in FIG. 10 , the causal graph acquisition unit 5 acquires a plurality of causal graphs G showing different causal relationships between a plurality of types of attributes and treatment effects. FIG. 10 is an explanatory diagram showing a causal graph when the plurality of types of attributes are age, gender, and cholesterol level. FIG. 10A shows a causal graph when the attribute directly related to the infection rate is gender, and FIG. 10B shows a causal graph when the attribute directly related to the infection rate is age. FIG. 10C shows a causal graph when the attributes directly related to the infection rate are both gender and age, and FIG. 10D shows a causal graph when the attribute directly related to the infection rate is cholesterol level. The causal graph acquisition unit 5 outputs the plurality of causal graphs to the likelihood calculation unit 6 and the causal graph selection unit 7, respectively.
[0073] The likelihood calculation unit 6 receives the observed data D ０ and each type of attribute X Ｉ，ｉ The first observation y １ Ｉ，ｉ The likelihood calculation unit 6 acquires a plurality of causal graphs G from the causal graph acquisition unit 5. The likelihood calculation unit 6 acquires the observed data D ０ and the first observation y １ Ｉ，ｉ and each causal graph g ｊ The likelihood P(D|g ｊ The likelihood calculation unit 6 calculates the likelihood of each causal graph g ｊ The likelihood P(D|g ｊ ) to the causal graph selection unit 7. The likelihood calculation process by the likelihood calculation unit 6 will now be described in detail.
[0074] First, the likelihood calculation unit 6 calculates the observed data D ０ and multiple types of attribute X Ｉ，ｉ The first observation y １ Ｉ，ｉ Observation data D １ and the observation data after integration is designated as D.
[0075]
[0076] Next, the likelihood calculation unit 6 calculates each causal graph g using the integrated observation data D as shown in the following equations (7) to (9). ｊ The likelihood P(D|g ｊ ) is calculated.
[0077] In formula (7), g ｊ is the causal graph, θ ｊ is g ｊ parameter, X ｊ is the causal graph g ｊ There is a set of Xs that are parents of Y.
[0078] The causal graph selection unit 7 acquires a plurality of causal graphs G from the causal graph acquisition unit 5, and selects each of the causal graphs g from the likelihood calculation unit 6. ｊ The likelihood P(D|g ｊ The causal graph selection unit 7 obtains each causal graph g ｊ The likelihood P(D|g ｊ ) based on the causal graph G, ｊ The causal graph selection unit 7 selects the selected causal graph g ｊ is output to the display processing unit 8. The process of selecting a causal graph by the causal graph selection unit 7 will now be described in detail.
[0079] First, the causal graph selection unit 7 selects each causal graph g according to Bayes' theorem as shown in the following equations (10) and (11). ｊ The likelihood P(D|g ｊ ), the probability P(g ｊ |D) is calculated.
[0080] In formula (11), N ｇ is the causal graph g ｊ is the number of
[0081] The causal graph selection unit 7 selects a causal graph G from among a plurality of causal graphs G with a probability P(g ｊ |D) is the largest causal graph g ｊ The causal graph g shown in FIGS. ｊ For example, the causal graph g shown in FIG.ｊ The probability P(g ｊ |D) is another causal graph g ｊ The probability P(g ｊ |D), the causal graph selection unit 7 selects the causal graph g ｊ Select .
[0082] The display processing unit 8 displays the selected causal graph g from the causal graph selection unit 7. ｊ The display processing unit 8 acquires the selected causal graph g ｊ is displayed on a display (not shown), for example.
[0083] In the above-described fourth embodiment, the optimal attribute search device is configured to include a causal graph acquisition unit 5 that acquires multiple causal graphs showing different causal relationships between multiple types of attributes and treatment effects, a likelihood calculation unit 6 that calculates the likelihood of each causal graph acquired by the causal graph acquisition unit 5 using first observation values related to each attribute, and a causal graph selection unit 7 that selects one of the multiple causal graphs based on the likelihood of the causal graph calculated by the likelihood calculation unit 6. Therefore, when multiple values of an attribute given as a variable are restricted to fall within a certain range, even if the optimal value of the attribute is outside the certain range, the optimal attribute search device can suppress deterioration in search accuracy for the optimal value and can present the causal relationships between multiple types of attributes and treatment effects.
[0084] In addition, the present disclosure allows for free combination of the respective embodiments, modification of any of the components of the respective embodiments, or omission of any of the components of the respective embodiments.
[0085] The present disclosure can suppress deterioration in the search accuracy for optimal values even when multiple values in an attribute given as a variable are restricted to fall within a certain range and the optimal value of the attribute is outside the certain range, and can be used in an optimal attribute search device, an optimal attribute search method, and a program.
[0086] 1 Data storage unit, 2 Optimum value search unit, 3 Observation value acquisition unit, 4 Function update unit, 5 Causal graph acquisition unit, 6 Likelihood calculation unit, 7 Causal graph selection unit, 8 Display processing unit, 11 Data storage circuit, 12 Optimum value search circuit, 13 Observation value acquisition circuit, 14 Function update circuit, 15 Causal graph acquisition circuit, 16 Likelihood calculation circuit, 17 Causal graph selection circuit, 18 Display processing circuit, 21 Memory, 22 Processor.
Claims
1. Given that each of several values in an attribute of a certain object is given as a variable, a first acquisition function is obtained that outputs the difference between the effect of the treatment when a certain treatment is applied to the object and the effect of the treatment when the treatment is not applied to the object. Based on the output of the first acquisition function, an optimal value search unit searches for one of the several values in the attribute to be the first optimal value. An observation value acquisition unit acquires the observed value of the treatment that corresponds to the first optimal value found by the optimal value search unit as the first observed value, The system includes a function update unit that updates the first acquisition function using the first observed values acquired by the observed value acquisition unit, The aforementioned optimal value search unit, Based on the output of the first acquisition function after the update by the function update unit, the first optimal value is searched for and the searched first optimal value is output. An optimal attribute search device characterized by the following features.
2. The aforementioned optimal value search unit, In addition to the first acquisition function described above, Given that each of the multiple values in the aforementioned target attribute is given as a variable, a second acquisition function is obtained that outputs the variance of the treatment's effectiveness, and based on the output of the second acquisition function, one of the multiple values in the attribute is searched for as the second optimal value. The aforementioned observation value acquisition unit is: Instead of obtaining the first observed value, The observed value of the treatment's effectiveness corresponding to the second optimal value found by the optimal value search unit is acquired as the second observed value. The function update unit, The first acquisition function is updated using the second observation value acquired by the observation value acquisition unit, instead of the first observation value. The optimal attribute search device according to claim 1, characterized in that it is a feature of the device described in claim 1.
3. The aforementioned optimal value search unit, In addition to the first acquisition function described above, Given that each of the multiple values in the aforementioned target attribute is given as a variable, a second acquisition function is obtained that outputs the variance of the treatment's effectiveness, and based on the output of the second acquisition function, one of the multiple values in the attribute is searched for as the second optimal value. The aforementioned observation value acquisition unit is: In addition to obtaining the first observed value, The observed value of the treatment's effectiveness corresponding to the second optimal value found by the optimal value search unit is acquired as the second observed value. The function update unit, The first acquisition function is updated using the first and second observed values. The optimal attribute search device according to claim 1, characterized in that it is a feature of the device described in claim 1.
4. The optimal value search process by the optimal value search unit, the observation value acquisition process by the observation value acquisition unit, and the update process of the first acquisition function by the function update unit are repeated until the difference output from the first acquisition function is equal to or greater than the target value. The optimal attribute search device according to any one of claims 1 to 3.
5. The aforementioned optimal value search unit, Based on the output of the first acquisition function after the update by the function update unit, the first optimal value is searched for and the searched first optimal value is output, as well as the difference output from the updated first acquisition function. The optimal attribute search device according to any one of claims 1 to 3.
6. There are multiple types of attributes of the aforementioned target, A causal graph that shows the causal relationship between the aforementioned multiple types of attributes and the effectiveness of the treatment, comprising a causal graph acquisition unit that acquires multiple causal graphs that show different causal relationships from one another, A likelihood calculation unit calculates the likelihood of each causal graph obtained by the causal graph acquisition unit using the first observed values relating to each type of attribute, A causal graph selection unit selects one of the multiple causal graphs based on the likelihood of the causal graph calculated by the likelihood calculation unit. The optimal attribute search device according to any one of claims 1 to 3, characterized by comprising:
7. Display processing unit that displays the causal graph selected by the aforementioned causal graph selection unit. The optimal attribute search device according to claim 6, characterized by comprising the following:
8. The optimal value search unit, given multiple values for an attribute of a certain object as variables, obtains a first acquisition function that outputs the difference between the effect of a treatment being applied to the object and the effect of not applying the treatment being applied to the object, and based on the output of the first acquisition function, searches for one of the multiple values for the attribute to be the first optimal value. The observation value acquisition unit acquires the observed value of the treatment that corresponds to the first optimal value found by the optimal value search unit as the first observed value. The function update unit updates the first acquisition function using the first observed value acquired by the observed value acquisition unit, The aforementioned optimal value search unit, Based on the output of the first acquisition function after the update by the function update unit, the first optimal value is searched for and the searched first optimal value is output. A method for finding optimal attributes, characterized by the following features.
9. A program to be executed by the computer of the optimal attribute search device, To the aforementioned computer, Given that multiple values for an attribute of a certain object are each given as variables, a first acquisition function is obtained that outputs the difference between the effect of the treatment when a certain treatment is applied to the object and the effect of the treatment when the treatment is not applied to the object. Based on the output of the first acquisition function, a search process is performed to find one of the multiple values for the attribute as the first optimal value. An acquisition process to acquire the observed value of the treatment that corresponds to the first optimal value as the first observed value, After performing an update process that updates the first acquisition function using the first observed value, A program for performing output processing that searches for the first optimal value based on the output of the first acquisition function after the update, and outputs the searched first optimal value.